Research And Application Of Species Distribution Model Based On Neural Network And Gray Wolf Optimization Algorithm

Neural networks use back propagation methods to simulate the distribution of species and are particularly suitable for solving information processing processes that are difficult to express by conventional methods.However,because of its complex structure,there is uncertainty in parameter setting,which can affect the accuracy of model simulation.In this paper,we first used the good global search and fast convergence ability of the gray wolf algorithm to improve the shortcomings of the neural network,which is easy to fall into local optimum and slow convergence.And then we obtained the gray wolf optimization neural network algorithm(GNNA)model,and compared with the commonly used species distribution models(i.e.ANN,GAM,GLM,Max Ent and RF)though three statistical test indexes(i.e.AUC,TSS and KAPPA).In addition,we established the distribution models using six species distribution models for an important medicinal plant of the Tibetan Plateau,Rhodiola chrysanthemifolia,and analyzed the goodness of simulation results of different models.Then we predicted the potential geographic distribution of R.chrysanthemifolia under present and future climate.The visualization of the potential geographic distribution and suitable classes of R.chrysanthemifoliawas realized by Arc GIS.The main conclusions drawn in this paper.1.GNNA model showed excellent prediction effect in simulating species distribution,with significant improvement in prediction accuracy compared with ANN,and was competitive with three species distribution models with better prediction effect(i.e.RF,GLM and Max Ent),especially the best performance of GNNA in predicting species distribution of global samples.2.In the conservation of R.chrysanthemifolia,the GNNA model also predicted the potential distribution significantly better than the ANN model,and its AUC,TSS and KAPPA values reached 0.93,0.79 and 0.78,respectively,with very good prediction results.In addition,the GNNA model was better than the GLM,GAM and RF models in predicting the potential distribution of R.chrysanthemifolia,except for the slightly lower prediction ability thanthe Max Ent model.3.The potential geographic distribution of R.chrysanthemifolia predicted by different modeling algorithms varied significantly,and the consistent results of the six models indicated that the current high suitable areas of R.chrysanthemifolia were mainly located in the eastern Himalayas and south-central Hengduan Mountains,and the suitableareas of R.chrysanthemifolia decreased under future climate change.Therefore,the combined results of the predictions of multiple species distribution models need to be considered when applying species distribution models for species conservation studies.